
302 Current Trends in Bayesian Methodology with Applications
that variation across runs at the end of the first pass is larger than numerical
standard error suggests (E (θ
′
1
x | Data) fo r runs A a nd B).
These suggestions could be investigated more critically with scores or hun-
dreds of runs of the SPS algorithm, but we conjecture the returns would be
low and in any event there is no basis for extrapolating results across models
and datasets. Most important, one cannot resort to this tac tic routinely in ap-
plied work. The results here support the earlier recommendation (at the end
of subsec tion 14.2.4) that the investigator proceed mainly using